Open Knowledge Format (OKF) is an open specification from Google Cloud, published June 12, 2026, that stores company knowledge as markdown files with YAML frontmatter so AI agents can read them without translation. A bundle is just a folder: no SDK, no database, no proprietary runtime. Every file needs exactly one required field, type. Everything else, links, sources, metrics, playbooks, is up to you.
Here is the contrarian part: OKF is not a new idea. It is Notion, Obsidian, and every internal wiki your team already abandoned, standardized enough that an agent can finally trust it. The technology was never the bottleneck. The discipline was.
Nintendo built a fortress. OKF tears one down.
Nintendo never wins on raw specs. The Switch shipped with a mobile-class chip while Sony and Microsoft chased teraflops, and Nintendo won anyway, because it controlled the ecosystem the hardware lived in: first-party games, proprietary formats, no cross-play, no compromise. Lock-in was the strategy.
OKF does the opposite on purpose. Markdown files. YAML frontmatter. No registry, no vendor SDK, no account required to read or write a bundle. Google is publishing a format that actively refuses to lock anyone into Google Cloud. That is unusual for a company that owns one of the three major model providers, and it is the detail worth sitting with before anything else in this article.
What Open Knowledge Format actually is
Open Knowledge Format is a directory-based spec that represents a « concept », a table, a playbook, a metric, an API, as one markdown file with a small YAML frontmatter block on top.
A bundle looks like this:
sales/
├── index.md
├── datasets/orders_db.md
├── tables/orders.md
└── metrics/weekly_active_users.mdEach file carries structured fields (type, title, description, resource, tags, timestamp) and a markdown body underneath. Concepts link to each other with ordinary markdown links, which turns the folder into a graph an agent can traverse instead of a pile of documents it has to guess about.
Three design choices carry the whole spec. First, it is minimally opinionated: type is the only mandatory field, so teams keep control of their content model. Second, producer and consumer are independent: a human writes the file, an agent reads it, a different agent updates it, and nobody needs to agree on tooling. Third, it is a format, not a platform: no cloud account required, ever, per Google’s own launch post (Google Cloud, June 2026).
AI researcher Andrej Karpathy named the underlying pattern in an April 2026 gist he called the « LLM wiki »: models don’t get bored, don’t forget to update a cross-reference, and can touch fifteen files in one pass, exactly the bookkeeping that makes humans quit maintaining personal wikis. OKF formalizes that pattern into something interoperable across teams and tools.
RAG finds fragments. OKF hands over concepts.
A retrieval system answers « what does this say » by pulling the nearest-matching chunks out of a vector store. It has no idea whether the chunk it just retrieved is current, contradicted by a newer decision, or missing the source that justifies it.
Retrieval-augmented generation depends entirely on the quality of what gets ingested, per Google Cloud’s own comparison table published alongside the spec. RAG stores embeddings; OKF stores curated, cross-linked concepts an agent reads and updates directly. The difference shows up the moment two documents disagree, RAG returns both fragments and lets the model guess; a well-linked OKF bundle points to which one is current and why.
Enterprises are already voting with their budgets. Buyer intent for hybrid retrieval architecture tripled from 10.3% to 33.3% in a single quarter in early 2026, per VentureBeat’s May 2026 reporting, as companies discovered that RAG alone could not carry production-grade agentic workloads. Retrieval optimization overtook evaluation as the top enterprise AI investment priority for the first time.
RAG and structured context are not rivals fighting for the same job. RAG is a search mechanism over a corpus. OKF is what makes that corpus worth searching in the first place.
The market already moved before most companies noticed
Context engineering, designing the informational environment an agent operates in, as distinct from prompt engineering, which optimizes the question, is now treated as its own discipline rather than a subset of prompting, according to 2026 industry analysis from Mem0 and others.
The gap between ambition and reality is the real story. Almost four in five enterprises have adopted AI agents in some form, yet only one in nine runs them in true production, the largest deployment backlog in enterprise technology history, per 2026 enterprise AI agent adoption research. Separately, 60% of large enterprises report agents already in production-level deployments, and 88% of executives plan to increase AI budgets specifically because of agentic initiatives.
The two figures are not contradictory; they describe different rungs of the same ladder, and most companies are stuck on a lower one than they think. The bottleneck is not the model. It’s the mess of docs, tribal knowledge, and dead wikis sitting behind it.
Quality, not capability, is what’s blocking the climb. 32% of practitioners cite output quality as their top barrier to production agents, per LangChain’s 2026 survey of over 1,300 practitioners, and that number sits right next to a governance figure that should worry every executive treating agents as a checkbox: only 21% of organizations have mature governance for agentic AI, despite 74% expecting to use agents moderately or more by 2027, per Deloitte’s 2026 State of AI report.
Five things worth structuring first
Sales relance. Open devis, objections, CRM exports, and client proof points become one playbook file with sources attached. An agent drafts a follow-up that is sourced and account-specific instead of generic, per the OKF use-case examples for commercial teams.
Marketing and communication. Positioning, validated messaging, and an editorial calendar become concepts an agent can reuse without drifting off-brand every third post.
RFP response. References, past offers, and prior answers become linked concepts an agent assembles into a first draft, cutting the research time that usually eats the first two days of any bid.
Support and operations. FAQs, recurring tickets, and internal procedures become a bundle an agent consults before answering, which is the difference between a response that’s merely fast and one that’s actually correct.
Incident runbooks. Each runbook is one file. An on-call agent reads the index, follows the links, and resolves a join path or a rollback step without paging a human at 3 a.m., per Google’s reference use cases.
Notice what’s missing from that list: nothing here requires new software. It requires someone to sit down and turn what already exists in someone’s head into a file with a type field.
| Approach | Storage | Schema required | Portable | SDK needed | Agent-readable |
|---|---|---|---|---|---|
| OKF v0.1 | Markdown + YAML | Only type | Yes | No | Yes, no translation |
| Notion | Proprietary DB | Per-workspace | Export-only | API needed | Via API only |
| Obsidian vault | Markdown files | None enforced | Yes | No | Bespoke conventions |
| Metadata catalog | Vendor store | Vendor schema | Export-only | Vendor SDK | Vendor-specific |
| RAG index alone | Vector store | Embedding model | No | Yes | Chunks, not concepts |
Source: adapted from Google Cloud’s OKF comparison, June 2026.
Own your context. Don’t rent it.
Every SaaS wiki you’ve ever used made the same trade: convenience today, lock-in tomorrow. Notion holds your knowledge in a proprietary database you can only leave through an export button designed to be as painful as legal allows. When the AI layer gets built on top of that wiki, you’ve rented your own company’s memory back from a vendor.
An OKF bundle is a folder of text files under version control. You can move it, diff it, restore last month’s version, or hand it to a completely different agent framework tomorrow without asking anyone’s permission. That’s not a technical nuance — it’s the entire strategic argument, and it echoes a truth from a very different kind of machine: the systems that last are the ones you own outright, not the ones you’re renting access to.
Foundation models are becoming a commodity layer, interchangeable, priced down, accessible to anyone with an API key. When everyone can call the same models, the differentiation stops being « which AI do you use » and starts being « what does your AI actually know. » Context is the moat now. A format you own is the only kind of moat that survives a vendor going out of business, changing its pricing, or getting acquired.
Before you build a bundle, ask what most companies get wrong here: is your context worth structuring, or does it not exist in usable form yet? Most SMBs answer that question honestly and discover the real first project isn’t OKF. It’s simply writing down decisions and playbooks that currently live only in someone’s head. OKF becomes valuable the moment that knowledge exists in any written form; it’s worthless if the knowledge was never written down at all.
Agents don’t need a smarter model next quarter. They need the runbook, the pricing logic, and the objection-handling playbook that’s currently trapped in a Slack thread from March. Write that down as a typed markdown file with a source, and every agent you plug in afterward starts smarter than the last one.
FAQ
Q: What is Open Knowledge Format (OKF)?
A: OKF is an open specification from Google Cloud, published June 12, 2026, that represents company knowledge as a directory of markdown files with YAML frontmatter, designed to be readable by humans and parseable by AI agents without a translation layer.
Q: Is OKF a Google product I need to sign up for?
A: No. OKF is a format, not a platform. It requires no cloud account, no SDK, and no proprietary runtime to read, write, or serve — Google Cloud has stated this explicitly and integrated it into its own Knowledge Catalog as one implementation among many possible ones.
Q: How is OKF different from RAG?
A: RAG retrieves fragments from a vector store at query time and depends heavily on the quality of ingested documents. OKF stores curated, cross-linked concepts with sources and types that an agent can navigate directly. Most production systems in 2026 use both together rather than choosing one.
Q: Does OKF replace Notion or Obsidian?
A: Not necessarily. OKF can sit as a portable layer exported from tools you already use, without forcing a full migration, according to Google’s own FAQ on the spec.
Q: Is OKF only useful for large enterprises with data teams?
A: No. A solo founder can use it as a structured second brain for offers, decisions, and notes; a small team can use it to standardize client relance and support answers. The format scales down as easily as it scales up.
Q: Is structuring context actually worth the time for a small company?
A: Most consultants will tell you yes reflexively. The honest answer is conditional: it’s worth it only if the underlying knowledge already exists somewhere, even messily. If nothing is written down yet, the first project is writing it down, OKF is the second step, not the first.
Q: What should I structure first if I’m starting from zero?
A: Whatever your team gets asked to repeat most often, a sales objection, a support answer, an onboarding step. Structure the highest-frequency question before the most complex one.
Verdict
OKF is not the interesting news. The interesting news is that Google just admitted, in public, that the knowledge sitting in your team’s heads is worth more than the model you’re calling, and handed you a format to prove it, for free, with no strings attached. Companies that spend the next quarter writing down what they already know will out-execute companies that spend it shopping for a better model.
If you’re structuring internal knowledge for AI agents and want a second set of eyes on the first bundle, Asymmetriq picks up exactly where a DIY OKF folder tends to stall: turning scattered playbooks into something an agent can actually act on, instead of a $4,700-a-month SaaS subscription renting your own knowledge back to you.